https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114767
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
Last reconfirmed| |2024-04-18
Blocks| |53947
Component|fortran |tree-optimization
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
When you avoid -ffast-math you'll get
.L4:
vmovupd (%rax), %ymm0
addq $32, %rax
vmulpd %ymm2, %ymm0, %ymm1
vpermilpd $5, %ymm0, %ymm0
vaddsubpd %ymm0, %ymm1, %ymm1
vmovupd %ymm1, -32(%rax)
cmpq %rax, %rdx
jne .L4
this consumes the negation with the vaddsubpd but has a multiplication
by zero for the sake of preserving signed zeros.
The main issue in the way of optimal code is that we lower the operations
early, arriving at
_6 = REALPART_EXPR <(*c_10(D))[_1]>;
_5 = IMAGPART_EXPR <(*c_10(D))[_1]>;
_14 = -_5;
REALPART_EXPR <(*c_10(D))[_1]> = _14;
IMAGPART_EXPR <(*c_10(D))[_1]> = _6;
and vectorizing that fails to use SLP which introduces an interleaving chain.
Negation isn't supported as "two-operators" op as there's no corresponding
"no negation" operation and we lack a way to insert a noop during SLP build.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations
